e44200579c8d53520b45485e436fd8dc24727ba1
[linux-2.6.git] / fs / nfs / direct.c
1 /*
2  * linux/fs/nfs/direct.c
3  *
4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
5  *
6  * High-performance uncached I/O for the Linux NFS client
7  *
8  * There are important applications whose performance or correctness
9  * depends on uncached access to file data.  Database clusters
10  * (multiple copies of the same instance running on separate hosts)
11  * implement their own cache coherency protocol that subsumes file
12  * system cache protocols.  Applications that process datasets
13  * considerably larger than the client's memory do not always benefit
14  * from a local cache.  A streaming video server, for instance, has no
15  * need to cache the contents of a file.
16  *
17  * When an application requests uncached I/O, all read and write requests
18  * are made directly to the server; data stored or fetched via these
19  * requests is not cached in the Linux page cache.  The client does not
20  * correct unaligned requests from applications.  All requested bytes are
21  * held on permanent storage before a direct write system call returns to
22  * an application.
23  *
24  * Solaris implements an uncached I/O facility called directio() that
25  * is used for backups and sequential I/O to very large files.  Solaris
26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
27  * an undocumented mount option.
28  *
29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
30  * help from Andrew Morton.
31  *
32  * 18 Dec 2001  Initial implementation for 2.4  --cel
33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
34  * 08 Jun 2003  Port to 2.5 APIs  --cel
35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
36  * 15 Sep 2004  Parallel async reads  --cel
37  * 04 May 2005  support O_DIRECT with aio  --cel
38  *
39  */
40
41 #include <linux/errno.h>
42 #include <linux/sched.h>
43 #include <linux/kernel.h>
44 #include <linux/file.h>
45 #include <linux/pagemap.h>
46 #include <linux/kref.h>
47
48 #include <linux/nfs_fs.h>
49 #include <linux/nfs_page.h>
50 #include <linux/sunrpc/clnt.h>
51
52 #include <asm/system.h>
53 #include <asm/uaccess.h>
54 #include <asm/atomic.h>
55
56 #include "internal.h"
57 #include "iostat.h"
58
59 #define NFSDBG_FACILITY         NFSDBG_VFS
60
61 static struct kmem_cache *nfs_direct_cachep;
62
63 /*
64  * This represents a set of asynchronous requests that we're waiting on
65  */
66 struct nfs_direct_req {
67         struct kref             kref;           /* release manager */
68
69         /* I/O parameters */
70         struct nfs_open_context *ctx;           /* file open context info */
71         struct kiocb *          iocb;           /* controlling i/o request */
72         struct inode *          inode;          /* target file of i/o */
73
74         /* completion state */
75         atomic_t                io_count;       /* i/os we're waiting for */
76         spinlock_t              lock;           /* protect completion state */
77         ssize_t                 count,          /* bytes actually processed */
78                                 error;          /* any reported error */
79         struct completion       completion;     /* wait for i/o completion */
80
81         /* commit state */
82         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
83         struct nfs_write_data * commit_data;    /* special write_data for commits */
84         int                     flags;
85 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
86 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
87         struct nfs_writeverf    verf;           /* unstable write verifier */
88 };
89
90 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
91 static const struct rpc_call_ops nfs_write_direct_ops;
92
93 static inline void get_dreq(struct nfs_direct_req *dreq)
94 {
95         atomic_inc(&dreq->io_count);
96 }
97
98 static inline int put_dreq(struct nfs_direct_req *dreq)
99 {
100         return atomic_dec_and_test(&dreq->io_count);
101 }
102
103 /**
104  * nfs_direct_IO - NFS address space operation for direct I/O
105  * @rw: direction (read or write)
106  * @iocb: target I/O control block
107  * @iov: array of vectors that define I/O buffer
108  * @pos: offset in file to begin the operation
109  * @nr_segs: size of iovec array
110  *
111  * The presence of this routine in the address space ops vector means
112  * the NFS client supports direct I/O.  However, we shunt off direct
113  * read and write requests before the VFS gets them, so this method
114  * should never be called.
115  */
116 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
117 {
118         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
119                         iocb->ki_filp->f_path.dentry->d_name.name,
120                         (long long) pos, nr_segs);
121
122         return -EINVAL;
123 }
124
125 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
126 {
127         unsigned int npages;
128         unsigned int i;
129
130         if (count == 0)
131                 return;
132         pages += (pgbase >> PAGE_SHIFT);
133         npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
134         for (i = 0; i < npages; i++) {
135                 struct page *page = pages[i];
136                 if (!PageCompound(page))
137                         set_page_dirty(page);
138         }
139 }
140
141 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
142 {
143         unsigned int i;
144         for (i = 0; i < npages; i++)
145                 page_cache_release(pages[i]);
146 }
147
148 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
149 {
150         struct nfs_direct_req *dreq;
151
152         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
153         if (!dreq)
154                 return NULL;
155
156         kref_init(&dreq->kref);
157         kref_get(&dreq->kref);
158         init_completion(&dreq->completion);
159         INIT_LIST_HEAD(&dreq->rewrite_list);
160         dreq->iocb = NULL;
161         dreq->ctx = NULL;
162         spin_lock_init(&dreq->lock);
163         atomic_set(&dreq->io_count, 0);
164         dreq->count = 0;
165         dreq->error = 0;
166         dreq->flags = 0;
167
168         return dreq;
169 }
170
171 static void nfs_direct_req_free(struct kref *kref)
172 {
173         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
174
175         if (dreq->ctx != NULL)
176                 put_nfs_open_context(dreq->ctx);
177         kmem_cache_free(nfs_direct_cachep, dreq);
178 }
179
180 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
181 {
182         kref_put(&dreq->kref, nfs_direct_req_free);
183 }
184
185 /*
186  * Collects and returns the final error value/byte-count.
187  */
188 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
189 {
190         ssize_t result = -EIOCBQUEUED;
191
192         /* Async requests don't wait here */
193         if (dreq->iocb)
194                 goto out;
195
196         result = wait_for_completion_killable(&dreq->completion);
197
198         if (!result)
199                 result = dreq->error;
200         if (!result)
201                 result = dreq->count;
202
203 out:
204         return (ssize_t) result;
205 }
206
207 /*
208  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
209  * the iocb is still valid here if this is a synchronous request.
210  */
211 static void nfs_direct_complete(struct nfs_direct_req *dreq)
212 {
213         if (dreq->iocb) {
214                 long res = (long) dreq->error;
215                 if (!res)
216                         res = (long) dreq->count;
217                 aio_complete(dreq->iocb, res, 0);
218         }
219         complete_all(&dreq->completion);
220
221         nfs_direct_req_release(dreq);
222 }
223
224 /*
225  * We must hold a reference to all the pages in this direct read request
226  * until the RPCs complete.  This could be long *after* we are woken up in
227  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
228  */
229 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
230 {
231         struct nfs_read_data *data = calldata;
232         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
233
234         if (nfs_readpage_result(task, data) != 0)
235                 return;
236
237         spin_lock(&dreq->lock);
238         if (unlikely(task->tk_status < 0)) {
239                 dreq->error = task->tk_status;
240                 spin_unlock(&dreq->lock);
241         } else {
242                 dreq->count += data->res.count;
243                 spin_unlock(&dreq->lock);
244                 nfs_direct_dirty_pages(data->pagevec,
245                                 data->args.pgbase,
246                                 data->res.count);
247         }
248         nfs_direct_release_pages(data->pagevec, data->npages);
249
250         if (put_dreq(dreq))
251                 nfs_direct_complete(dreq);
252 }
253
254 static const struct rpc_call_ops nfs_read_direct_ops = {
255         .rpc_call_done = nfs_direct_read_result,
256         .rpc_release = nfs_readdata_release,
257 };
258
259 /*
260  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
261  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
262  * bail and stop sending more reads.  Read length accounting is
263  * handled automatically by nfs_direct_read_result().  Otherwise, if
264  * no requests have been sent, just return an error.
265  */
266 static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
267                                                 const struct iovec *iov,
268                                                 loff_t pos)
269 {
270         struct nfs_open_context *ctx = dreq->ctx;
271         struct inode *inode = ctx->path.dentry->d_inode;
272         unsigned long user_addr = (unsigned long)iov->iov_base;
273         size_t count = iov->iov_len;
274         size_t rsize = NFS_SERVER(inode)->rsize;
275         struct rpc_task *task;
276         struct rpc_message msg = {
277                 .rpc_cred = ctx->cred,
278         };
279         struct rpc_task_setup task_setup_data = {
280                 .rpc_client = NFS_CLIENT(inode),
281                 .rpc_message = &msg,
282                 .callback_ops = &nfs_read_direct_ops,
283                 .workqueue = nfsiod_workqueue,
284                 .flags = RPC_TASK_ASYNC,
285         };
286         unsigned int pgbase;
287         int result;
288         ssize_t started = 0;
289
290         do {
291                 struct nfs_read_data *data;
292                 size_t bytes;
293
294                 pgbase = user_addr & ~PAGE_MASK;
295                 bytes = min(rsize,count);
296
297                 result = -ENOMEM;
298                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
299                 if (unlikely(!data))
300                         break;
301
302                 down_read(&current->mm->mmap_sem);
303                 result = get_user_pages(current, current->mm, user_addr,
304                                         data->npages, 1, 0, data->pagevec, NULL);
305                 up_read(&current->mm->mmap_sem);
306                 if (result < 0) {
307                         nfs_readdata_release(data);
308                         break;
309                 }
310                 if ((unsigned)result < data->npages) {
311                         bytes = result * PAGE_SIZE;
312                         if (bytes <= pgbase) {
313                                 nfs_direct_release_pages(data->pagevec, result);
314                                 nfs_readdata_release(data);
315                                 break;
316                         }
317                         bytes -= pgbase;
318                         data->npages = result;
319                 }
320
321                 get_dreq(dreq);
322
323                 data->req = (struct nfs_page *) dreq;
324                 data->inode = inode;
325                 data->cred = msg.rpc_cred;
326                 data->args.fh = NFS_FH(inode);
327                 data->args.context = get_nfs_open_context(ctx);
328                 data->args.offset = pos;
329                 data->args.pgbase = pgbase;
330                 data->args.pages = data->pagevec;
331                 data->args.count = bytes;
332                 data->res.fattr = &data->fattr;
333                 data->res.eof = 0;
334                 data->res.count = bytes;
335                 msg.rpc_argp = &data->args;
336                 msg.rpc_resp = &data->res;
337
338                 task_setup_data.task = &data->task;
339                 task_setup_data.callback_data = data;
340                 NFS_PROTO(inode)->read_setup(data, &msg);
341
342                 task = rpc_run_task(&task_setup_data);
343                 if (!IS_ERR(task))
344                         rpc_put_task(task);
345
346                 dprintk("NFS: %5u initiated direct read call "
347                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
348                                 data->task.tk_pid,
349                                 inode->i_sb->s_id,
350                                 (long long)NFS_FILEID(inode),
351                                 bytes,
352                                 (unsigned long long)data->args.offset);
353
354                 started += bytes;
355                 user_addr += bytes;
356                 pos += bytes;
357                 /* FIXME: Remove this unnecessary math from final patch */
358                 pgbase += bytes;
359                 pgbase &= ~PAGE_MASK;
360                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
361
362                 count -= bytes;
363         } while (count != 0);
364
365         if (started)
366                 return started;
367         return result < 0 ? (ssize_t) result : -EFAULT;
368 }
369
370 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
371                                               const struct iovec *iov,
372                                               unsigned long nr_segs,
373                                               loff_t pos)
374 {
375         ssize_t result = -EINVAL;
376         size_t requested_bytes = 0;
377         unsigned long seg;
378
379         get_dreq(dreq);
380
381         for (seg = 0; seg < nr_segs; seg++) {
382                 const struct iovec *vec = &iov[seg];
383                 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
384                 if (result < 0)
385                         break;
386                 requested_bytes += result;
387                 if ((size_t)result < vec->iov_len)
388                         break;
389                 pos += vec->iov_len;
390         }
391
392         if (put_dreq(dreq))
393                 nfs_direct_complete(dreq);
394
395         if (requested_bytes != 0)
396                 return 0;
397
398         if (result < 0)
399                 return result;
400         return -EIO;
401 }
402
403 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
404                                unsigned long nr_segs, loff_t pos)
405 {
406         ssize_t result = 0;
407         struct inode *inode = iocb->ki_filp->f_mapping->host;
408         struct nfs_direct_req *dreq;
409
410         dreq = nfs_direct_req_alloc();
411         if (!dreq)
412                 return -ENOMEM;
413
414         dreq->inode = inode;
415         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
416         if (!is_sync_kiocb(iocb))
417                 dreq->iocb = iocb;
418
419         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
420         if (!result)
421                 result = nfs_direct_wait(dreq);
422         nfs_direct_req_release(dreq);
423
424         return result;
425 }
426
427 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
428 {
429         while (!list_empty(&dreq->rewrite_list)) {
430                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
431                 list_del(&data->pages);
432                 nfs_direct_release_pages(data->pagevec, data->npages);
433                 nfs_writedata_release(data);
434         }
435 }
436
437 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
438 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
439 {
440         struct inode *inode = dreq->inode;
441         struct list_head *p;
442         struct nfs_write_data *data;
443         struct rpc_task *task;
444         struct rpc_message msg = {
445                 .rpc_cred = dreq->ctx->cred,
446         };
447         struct rpc_task_setup task_setup_data = {
448                 .rpc_client = NFS_CLIENT(inode),
449                 .callback_ops = &nfs_write_direct_ops,
450                 .workqueue = nfsiod_workqueue,
451                 .flags = RPC_TASK_ASYNC,
452         };
453
454         dreq->count = 0;
455         get_dreq(dreq);
456
457         list_for_each(p, &dreq->rewrite_list) {
458                 data = list_entry(p, struct nfs_write_data, pages);
459
460                 get_dreq(dreq);
461
462                 /* Use stable writes */
463                 data->args.stable = NFS_FILE_SYNC;
464
465                 /*
466                  * Reset data->res.
467                  */
468                 nfs_fattr_init(&data->fattr);
469                 data->res.count = data->args.count;
470                 memset(&data->verf, 0, sizeof(data->verf));
471
472                 /*
473                  * Reuse data->task; data->args should not have changed
474                  * since the original request was sent.
475                  */
476                 task_setup_data.task = &data->task;
477                 task_setup_data.callback_data = data;
478                 msg.rpc_argp = &data->args;
479                 msg.rpc_resp = &data->res;
480                 NFS_PROTO(inode)->write_setup(data, &msg);
481
482                 /*
483                  * We're called via an RPC callback, so BKL is already held.
484                  */
485                 task = rpc_run_task(&task_setup_data);
486                 if (!IS_ERR(task))
487                         rpc_put_task(task);
488
489                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
490                                 data->task.tk_pid,
491                                 inode->i_sb->s_id,
492                                 (long long)NFS_FILEID(inode),
493                                 data->args.count,
494                                 (unsigned long long)data->args.offset);
495         }
496
497         if (put_dreq(dreq))
498                 nfs_direct_write_complete(dreq, inode);
499 }
500
501 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
502 {
503         struct nfs_write_data *data = calldata;
504         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
505
506         /* Call the NFS version-specific code */
507         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
508                 return;
509         if (unlikely(task->tk_status < 0)) {
510                 dprintk("NFS: %5u commit failed with error %d.\n",
511                                 task->tk_pid, task->tk_status);
512                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
513         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
514                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
515                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
516         }
517
518         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
519         nfs_direct_write_complete(dreq, data->inode);
520 }
521
522 static const struct rpc_call_ops nfs_commit_direct_ops = {
523         .rpc_call_done = nfs_direct_commit_result,
524         .rpc_release = nfs_commit_release,
525 };
526
527 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
528 {
529         struct nfs_write_data *data = dreq->commit_data;
530         struct rpc_task *task;
531         struct rpc_message msg = {
532                 .rpc_argp = &data->args,
533                 .rpc_resp = &data->res,
534                 .rpc_cred = dreq->ctx->cred,
535         };
536         struct rpc_task_setup task_setup_data = {
537                 .task = &data->task,
538                 .rpc_client = NFS_CLIENT(dreq->inode),
539                 .rpc_message = &msg,
540                 .callback_ops = &nfs_commit_direct_ops,
541                 .callback_data = data,
542                 .workqueue = nfsiod_workqueue,
543                 .flags = RPC_TASK_ASYNC,
544         };
545
546         data->inode = dreq->inode;
547         data->cred = msg.rpc_cred;
548
549         data->args.fh = NFS_FH(data->inode);
550         data->args.offset = 0;
551         data->args.count = 0;
552         data->args.context = get_nfs_open_context(dreq->ctx);
553         data->res.count = 0;
554         data->res.fattr = &data->fattr;
555         data->res.verf = &data->verf;
556
557         NFS_PROTO(data->inode)->commit_setup(data, &msg);
558
559         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
560         dreq->commit_data = NULL;
561
562         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
563
564         task = rpc_run_task(&task_setup_data);
565         if (!IS_ERR(task))
566                 rpc_put_task(task);
567 }
568
569 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
570 {
571         int flags = dreq->flags;
572
573         dreq->flags = 0;
574         switch (flags) {
575                 case NFS_ODIRECT_DO_COMMIT:
576                         nfs_direct_commit_schedule(dreq);
577                         break;
578                 case NFS_ODIRECT_RESCHED_WRITES:
579                         nfs_direct_write_reschedule(dreq);
580                         break;
581                 default:
582                         if (dreq->commit_data != NULL)
583                                 nfs_commit_free(dreq->commit_data);
584                         nfs_direct_free_writedata(dreq);
585                         nfs_zap_mapping(inode, inode->i_mapping);
586                         nfs_direct_complete(dreq);
587         }
588 }
589
590 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
591 {
592         dreq->commit_data = nfs_commit_alloc();
593         if (dreq->commit_data != NULL)
594                 dreq->commit_data->req = (struct nfs_page *) dreq;
595 }
596 #else
597 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
598 {
599         dreq->commit_data = NULL;
600 }
601
602 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
603 {
604         nfs_direct_free_writedata(dreq);
605         nfs_zap_mapping(inode, inode->i_mapping);
606         nfs_direct_complete(dreq);
607 }
608 #endif
609
610 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
611 {
612         struct nfs_write_data *data = calldata;
613         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
614         int status = task->tk_status;
615
616         if (nfs_writeback_done(task, data) != 0)
617                 return;
618
619         spin_lock(&dreq->lock);
620
621         if (unlikely(status < 0)) {
622                 /* An error has occurred, so we should not commit */
623                 dreq->flags = 0;
624                 dreq->error = status;
625         }
626         if (unlikely(dreq->error != 0))
627                 goto out_unlock;
628
629         dreq->count += data->res.count;
630
631         if (data->res.verf->committed != NFS_FILE_SYNC) {
632                 switch (dreq->flags) {
633                         case 0:
634                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
635                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
636                                 break;
637                         case NFS_ODIRECT_DO_COMMIT:
638                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
639                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
640                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
641                                 }
642                 }
643         }
644 out_unlock:
645         spin_unlock(&dreq->lock);
646 }
647
648 /*
649  * NB: Return the value of the first error return code.  Subsequent
650  *     errors after the first one are ignored.
651  */
652 static void nfs_direct_write_release(void *calldata)
653 {
654         struct nfs_write_data *data = calldata;
655         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
656
657         if (put_dreq(dreq))
658                 nfs_direct_write_complete(dreq, data->inode);
659 }
660
661 static const struct rpc_call_ops nfs_write_direct_ops = {
662         .rpc_call_done = nfs_direct_write_result,
663         .rpc_release = nfs_direct_write_release,
664 };
665
666 /*
667  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
668  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
669  * bail and stop sending more writes.  Write length accounting is
670  * handled automatically by nfs_direct_write_result().  Otherwise, if
671  * no requests have been sent, just return an error.
672  */
673 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
674                                                  const struct iovec *iov,
675                                                  loff_t pos, int sync)
676 {
677         struct nfs_open_context *ctx = dreq->ctx;
678         struct inode *inode = ctx->path.dentry->d_inode;
679         unsigned long user_addr = (unsigned long)iov->iov_base;
680         size_t count = iov->iov_len;
681         struct rpc_task *task;
682         struct rpc_message msg = {
683                 .rpc_cred = ctx->cred,
684         };
685         struct rpc_task_setup task_setup_data = {
686                 .rpc_client = NFS_CLIENT(inode),
687                 .rpc_message = &msg,
688                 .callback_ops = &nfs_write_direct_ops,
689                 .workqueue = nfsiod_workqueue,
690                 .flags = RPC_TASK_ASYNC,
691         };
692         size_t wsize = NFS_SERVER(inode)->wsize;
693         unsigned int pgbase;
694         int result;
695         ssize_t started = 0;
696
697         do {
698                 struct nfs_write_data *data;
699                 size_t bytes;
700
701                 pgbase = user_addr & ~PAGE_MASK;
702                 bytes = min(wsize,count);
703
704                 result = -ENOMEM;
705                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
706                 if (unlikely(!data))
707                         break;
708
709                 down_read(&current->mm->mmap_sem);
710                 result = get_user_pages(current, current->mm, user_addr,
711                                         data->npages, 0, 0, data->pagevec, NULL);
712                 up_read(&current->mm->mmap_sem);
713                 if (result < 0) {
714                         nfs_writedata_release(data);
715                         break;
716                 }
717                 if ((unsigned)result < data->npages) {
718                         bytes = result * PAGE_SIZE;
719                         if (bytes <= pgbase) {
720                                 nfs_direct_release_pages(data->pagevec, result);
721                                 nfs_writedata_release(data);
722                                 break;
723                         }
724                         bytes -= pgbase;
725                         data->npages = result;
726                 }
727
728                 get_dreq(dreq);
729
730                 list_move_tail(&data->pages, &dreq->rewrite_list);
731
732                 data->req = (struct nfs_page *) dreq;
733                 data->inode = inode;
734                 data->cred = msg.rpc_cred;
735                 data->args.fh = NFS_FH(inode);
736                 data->args.context = get_nfs_open_context(ctx);
737                 data->args.offset = pos;
738                 data->args.pgbase = pgbase;
739                 data->args.pages = data->pagevec;
740                 data->args.count = bytes;
741                 data->args.stable = sync;
742                 data->res.fattr = &data->fattr;
743                 data->res.count = bytes;
744                 data->res.verf = &data->verf;
745
746                 task_setup_data.task = &data->task;
747                 task_setup_data.callback_data = data;
748                 msg.rpc_argp = &data->args;
749                 msg.rpc_resp = &data->res;
750                 NFS_PROTO(inode)->write_setup(data, &msg);
751
752                 task = rpc_run_task(&task_setup_data);
753                 if (!IS_ERR(task))
754                         rpc_put_task(task);
755
756                 dprintk("NFS: %5u initiated direct write call "
757                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
758                                 data->task.tk_pid,
759                                 inode->i_sb->s_id,
760                                 (long long)NFS_FILEID(inode),
761                                 bytes,
762                                 (unsigned long long)data->args.offset);
763
764                 started += bytes;
765                 user_addr += bytes;
766                 pos += bytes;
767
768                 /* FIXME: Remove this useless math from the final patch */
769                 pgbase += bytes;
770                 pgbase &= ~PAGE_MASK;
771                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
772
773                 count -= bytes;
774         } while (count != 0);
775
776         if (started)
777                 return started;
778         return result < 0 ? (ssize_t) result : -EFAULT;
779 }
780
781 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
782                                                const struct iovec *iov,
783                                                unsigned long nr_segs,
784                                                loff_t pos, int sync)
785 {
786         ssize_t result = 0;
787         size_t requested_bytes = 0;
788         unsigned long seg;
789
790         get_dreq(dreq);
791
792         for (seg = 0; seg < nr_segs; seg++) {
793                 const struct iovec *vec = &iov[seg];
794                 result = nfs_direct_write_schedule_segment(dreq, vec,
795                                                            pos, sync);
796                 if (result < 0)
797                         break;
798                 requested_bytes += result;
799                 if ((size_t)result < vec->iov_len)
800                         break;
801                 pos += vec->iov_len;
802         }
803
804         if (put_dreq(dreq))
805                 nfs_direct_write_complete(dreq, dreq->inode);
806
807         if (requested_bytes != 0)
808                 return 0;
809
810         if (result < 0)
811                 return result;
812         return -EIO;
813 }
814
815 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
816                                 unsigned long nr_segs, loff_t pos,
817                                 size_t count)
818 {
819         ssize_t result = 0;
820         struct inode *inode = iocb->ki_filp->f_mapping->host;
821         struct nfs_direct_req *dreq;
822         size_t wsize = NFS_SERVER(inode)->wsize;
823         int sync = NFS_UNSTABLE;
824
825         dreq = nfs_direct_req_alloc();
826         if (!dreq)
827                 return -ENOMEM;
828         nfs_alloc_commit_data(dreq);
829
830         if (dreq->commit_data == NULL || count < wsize)
831                 sync = NFS_FILE_SYNC;
832
833         dreq->inode = inode;
834         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
835         if (!is_sync_kiocb(iocb))
836                 dreq->iocb = iocb;
837
838         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
839         if (!result)
840                 result = nfs_direct_wait(dreq);
841         nfs_direct_req_release(dreq);
842
843         return result;
844 }
845
846 /**
847  * nfs_file_direct_read - file direct read operation for NFS files
848  * @iocb: target I/O control block
849  * @iov: vector of user buffers into which to read data
850  * @nr_segs: size of iov vector
851  * @pos: byte offset in file where reading starts
852  *
853  * We use this function for direct reads instead of calling
854  * generic_file_aio_read() in order to avoid gfar's check to see if
855  * the request starts before the end of the file.  For that check
856  * to work, we must generate a GETATTR before each direct read, and
857  * even then there is a window between the GETATTR and the subsequent
858  * READ where the file size could change.  Our preference is simply
859  * to do all reads the application wants, and the server will take
860  * care of managing the end of file boundary.
861  *
862  * This function also eliminates unnecessarily updating the file's
863  * atime locally, as the NFS server sets the file's atime, and this
864  * client must read the updated atime from the server back into its
865  * cache.
866  */
867 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
868                                 unsigned long nr_segs, loff_t pos)
869 {
870         ssize_t retval = -EINVAL;
871         struct file *file = iocb->ki_filp;
872         struct address_space *mapping = file->f_mapping;
873         size_t count;
874
875         count = iov_length(iov, nr_segs);
876         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
877
878         dprintk("nfs: direct read(%s/%s, %zd@%Ld)\n",
879                 file->f_path.dentry->d_parent->d_name.name,
880                 file->f_path.dentry->d_name.name,
881                 count, (long long) pos);
882
883         retval = 0;
884         if (!count)
885                 goto out;
886
887         retval = nfs_sync_mapping(mapping);
888         if (retval)
889                 goto out;
890
891         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
892         if (retval > 0)
893                 iocb->ki_pos = pos + retval;
894
895 out:
896         return retval;
897 }
898
899 /**
900  * nfs_file_direct_write - file direct write operation for NFS files
901  * @iocb: target I/O control block
902  * @iov: vector of user buffers from which to write data
903  * @nr_segs: size of iov vector
904  * @pos: byte offset in file where writing starts
905  *
906  * We use this function for direct writes instead of calling
907  * generic_file_aio_write() in order to avoid taking the inode
908  * semaphore and updating the i_size.  The NFS server will set
909  * the new i_size and this client must read the updated size
910  * back into its cache.  We let the server do generic write
911  * parameter checking and report problems.
912  *
913  * We also avoid an unnecessary invocation of generic_osync_inode(),
914  * as it is fairly meaningless to sync the metadata of an NFS file.
915  *
916  * We eliminate local atime updates, see direct read above.
917  *
918  * We avoid unnecessary page cache invalidations for normal cached
919  * readers of this file.
920  *
921  * Note that O_APPEND is not supported for NFS direct writes, as there
922  * is no atomic O_APPEND write facility in the NFS protocol.
923  */
924 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
925                                 unsigned long nr_segs, loff_t pos)
926 {
927         ssize_t retval = -EINVAL;
928         struct file *file = iocb->ki_filp;
929         struct address_space *mapping = file->f_mapping;
930         size_t count;
931
932         count = iov_length(iov, nr_segs);
933         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
934
935         dfprintk(VFS, "nfs: direct write(%s/%s, %zd@%Ld)\n",
936                 file->f_path.dentry->d_parent->d_name.name,
937                 file->f_path.dentry->d_name.name,
938                 count, (long long) pos);
939
940         retval = generic_write_checks(file, &pos, &count, 0);
941         if (retval)
942                 goto out;
943
944         retval = -EINVAL;
945         if ((ssize_t) count < 0)
946                 goto out;
947         retval = 0;
948         if (!count)
949                 goto out;
950
951         retval = nfs_sync_mapping(mapping);
952         if (retval)
953                 goto out;
954
955         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
956
957         if (retval > 0)
958                 iocb->ki_pos = pos + retval;
959
960 out:
961         return retval;
962 }
963
964 /**
965  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
966  *
967  */
968 int __init nfs_init_directcache(void)
969 {
970         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
971                                                 sizeof(struct nfs_direct_req),
972                                                 0, (SLAB_RECLAIM_ACCOUNT|
973                                                         SLAB_MEM_SPREAD),
974                                                 NULL);
975         if (nfs_direct_cachep == NULL)
976                 return -ENOMEM;
977
978         return 0;
979 }
980
981 /**
982  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
983  *
984  */
985 void nfs_destroy_directcache(void)
986 {
987         kmem_cache_destroy(nfs_direct_cachep);
988 }